Maximum Likelihood Sequence Detection

Transcription

1 1 The Channel Delay Spread Channel Model Matched Filter as Receiver Front End... 4 Detection Terms Maximum Lielihood Detection of a Single Symbol Maximum Lielihood Detection of a Signal Vector ML Detection with Intersymbol Interference Sequence Detection Marov Chains Marov Chain Signal Generator The Viterbi Algorithm Error Probability Calculation Error Event Symbol Error Probability Upper Bound of Detection Error Lower Bound of Detection Error Upper and Lower Bounds of Detection Error... 18

2 1 The Channel If electromagnetic energy carrying a modulated signal, propagates along more than one path connecting the transmitter to the receiver, this is called multipath propagation. Multipath propagation occurs when for example radio waves are reflected, scattered or refracted by e.g. buildings, cars or hills. 1.1 Delay Spread Signal components traveling along different paths suffer from different attenuations and delays, which combine at the receiver to produce a distorted version of the transmitted signal. To characterize the various delays incurred by the signal traveling through the channel the delay spread is introduced as the largest of the delays. The delay spread causes the two effects of time dispersion and frequency selective fading. In a time dispersion channel a signal bandwidth much larger than the inverse of the delay spread the various paths are distinguishable as they produce copies of the transmitted signal with different delays and attenuations. For narrower bandwidths however, the received signal copies tend to overlap. This generates a form of linear distortion nown as intersymbol interference (ISI). If the bandwidth is large, different frequencies of the signal typically suffer from different attenuations along different paths. This is called frequency selective fading which is not a tpic in this paper. 1. Channel Model Any further only passband PAM modulation and its degenerate case of baseband PAM are considered, since FSK is generally not used in situations where ISI is an important phenomenon. PAM is preferred in this situation because FSK is a nonlinear modulation technique and is therefore difficult to equalize properly. In general a channel can be modeled as a linear time-invariant filter with impulse response b(t) and additive noise N(t), as shown in Figure 1-1. Figure 1-1 Advanced Signal Processing WT 004 Page of 18

3 Since the input to this transmit filter is discrete-time, and the output is sampled by the sampler, we can replace the transmit filter, channel and receive filter plus the sampler with an equivalent discrete-time filter. With U = U( T) = [ N( t)* f( t )] t = T (1) and p = p( T) = [ g( t)* b( t)* f( t )] t = t. () The discrete-time equivalent channel model for PAM (as in Figure 1-1) is then shown in Figure 1-. A passband PAM system with transmit filter g(t) and receive filter f(t)can be modeled as a completely discrete-time system if the continuous-time subsystems are considered as part of the channel. The equivalent discrete-time channel has impulse response p and transfer function P(z). jωt π π π Pe ( ) = G j ω m BE j ω m F j ω m m= T T T (3) The equivalent noise Z = ZT ( ) has independent real and imaginary parts each with variance given by σ. The power spectrum of the noise is given by N ( ( π ω )) T 0 ( ω S T Z e ) = T m= F j + m where T equals the symbol rate. Figure 1- (4) Figure 1-3 Advanced Signal Processing WT 004 Page 3 of 18

4 1.3 Matched Filter as Receiver Front End The first part of the receiver the receive filter and the sampler are typically built of a matched filter that is sampled at symbol rate. The discrete-time equivalent channel model utilizing a sampled matched filter may be depicted as in Figure 1-4. Figure 1-4 Of fundamental importance to the channel model is the autocorrelation function for a given baseband receive pulse shape h(t) * h( ) = ( ) ( ) ρ h t h t T dt. (5) j T The Fourier transform of (5) autocorrelation S ( e ω ) is called folded spectrum: h jωt 1 ( ( π ρ ω )) (6) jωt Sh( e ) = h( ) e = H j + m T = T m= The folded spectrum is real valued, nonnegative and in general not symmetric. Now the discrete time channel that includes the matched filter is easily obtained using the discrete-time model for PAM systems. The receive filter from the discrete PAM model is replaced by the sampled matched * filter F( jω) = H ( jω). The impulse response of the baseband channel ρ h( ), the pulse j T autocorrelation function and the transfer function are equal to the folded spectrum S ( e ω ). Further, the equivalent discrete-time noise process 0 Z has power spectrum jωt jωt S ( e ) = N S ( e ). (7) Z h At this point the calculation of probability density functions of the received estimates P is laborious because the noise in this model is not white and therefore the random variables P are not independent. Therefore the colorization from the matched filter has to be compensated. This can be achieved by spectral factorization of S z = M z M z. (8) ( ) ( ) (1/ ) h γ Then Sh( z) = H ( z) Hap( z) can be decomposed to a imal phase filter and an allpass filter where H ( z) γ M( z) =. So adding the inverse imal phase filter to the matched filter compensates the transfer function to constant attenuation of all frequencies as in Figure 1-5. h Figure 1-5 Advanced Signal Processing WT 004 Page 4 of 18

5 0 Therefore a new channel model is obtained where ( ) N SN z = is the variance of the now K Gaussian noise process N and U is the received symbol distorted by ISI. Figure 1-6 Detection The general approach to deriving optimal receivers is to model the relationship between the transmitted and received signals by joint probability distributions. Based on a noisy observation (the received signal plus noise), we wish to estimate or detect the input signal. The term estimation is used when the transmitted signal is a continuous-valued random variable and the term detection when the transmitted signal is discrete-valued (even if the received signal is continuous-valued). So the first case refers to analog and the second to digital communication systems. In this section the following model will be used. Figure -1 The deteristic portion of the channel is called signal generation and the random component the noise generation. The input X is a discrete-time and discrete-valued random process and has only a finite number of possible values..1 Terms Alphabet A: the set of symbols those are available for transmission. A = {-1,+1,-j,+j} Symbol a, A : -th symbol from the alphabet. a 1 =-1 Size of the alphabet M: number of symbols in the alphabet. M = 4 Sample space of the Alphabet Ω A Symbol of the sample space â Advanced Signal Processing WT 004 Page 5 of 18

6 . Maximum Lielihood Detection of a Single Symbol In order to decide about the received data symbol, nowledge about the noise generator is needed. The noise generator is completely specified by the discrete distribution of Y conditioned in the nowledge of the data symbol pya ( y â ). The maximum lielihood (ML) detector chooses âε Ω A to maximize the lielihood pya ( y â ), where y is the observation of Y. The ML detection is a special case of maximum a-posteriori probability detection for the simplifying assumption that all the possible inputs p ( â) are equally liely. p AY ( â y) p ( y â) p ( â) A = YA A py ( y) (9) Example: Suppose that we have additive discrete noise N, so that Y = A+N. Assume A and N are independent and tae on values zero and one according to the flip of two fair coins. There are three possible observations, y = 0, 1, or. The lielihoods pya ( y â ) for the observation y=0 are: pya (0 0) = 0.5 p (0 1) = 0 YA So if the observation is y=0, the ML detector selects â = 0. If the observation is y=1, then the lielihoods are equal pya (1 0) = pya (1 1) = 0.5 and therefore the ML detector selects either zero or one. If the observation is y =, the ML detector selects â = 1. The probability of error Pr[ error] = Pr[ â a] is a measure of the quality of the detector. Example: From the previous example Y=A+N taes on values in {0, 1, } with probabilities {0.5, 0.5, 0.5}, respectively. If the observation is y = 0 or y = the ML detector does not mae errors. If the observation is y = 1, however, the detector is unsure and chooses for example at random. Then the detector will be wrong half the time. Since p Y (1) = 0.5, the total probability of error is Pr[ error ] = 0.5. Advanced Signal Processing WT 004 Page 6 of 18

7 .3 Maximum Lielihood Detection of a Signal Vector Since the goal in this paper is the detection of a sequence of symbols it is convenient to extend the ML detection of a single symbol to a vector of symbols. Here the signal generator accepts an input X and maps it into a vector signal S with dimension M. The observation is a vector Y with the same dimension as the signal. The noise generator f ( ) is specified by the conditional distribution of the observation given the signal YS y s. A common characteristic of the noise generator is independent noise components M f ( y s YS ) = f ( y s ) (10). Y S = 1 The ML detector chooses the signal vector ŝ from all the possibilities in order to maximize the conditional probability ˆ ( y s ), which is given directly from the noise generator. f YS For additive Gaussian noise with variance σ and independent noise components f ( y sˆ) = f ( y sˆ sˆ) = f ( y s ˆ) (11). YS NS N Figure - Hence the ML detector selects ŝ to maximize 1 1 f ( y ˆ ) exp ˆ s = N M / M ( π) σ σ y s (1) where M is the dimension of the vectors and f N is the PDF of the Noise. Maximizing this is equivalent to imizing y s ˆ, the Euclidian distance between the received vector y and the signal vector ŝ. Advanced Signal Processing WT 004 Page 7 of 18

8 .4 ML Detection with Intersymbol Interference Consider a situation where the input to the signal generator is a single data symbol A, and the signal generator is a linear time-invariant filter with impulse response h,0 M, assumed to be finite in extent. This filter represents the dispersion of the data symbol in time. A single isolated discrete-time impulse δ amplitude modulated by an isolated data symbol A is transmitted. The noise is independent additive and Gaussian generated. The observation Y is only considered within the range 0 M, since there are no signal components outside. Thus the observation may be written in vector notation as Y= ha + N. If the single data symbol A from alphabet Ω A, the ML detector selects âε Ω A to imize the Euclidian distance between the vectors âh and observation y. Thus the ML detector computes y hâ = y y, h â+ h â (13) and chooses â for which this is the imum. Since y does not depend on the decision, the ML detector equivalently selects â to maximize The term Figure -3 yh, â h â. (14) [ ] 0 yh, = yh m m = y * h (15) = m may be computed using a correlator or a matched filter as depicted in Figure -4. Example: Figure -4 Suppose that a symbol A from the alphabet { 0,1} system with impulse response h = δ + 0.5δ 1. Ω = is transmitted through the LTI A In this case M=1 and -5. h =1.5, so the ML detector can be implemented lie in Figure Advanced Signal Processing WT 004 Page 8 of 18

9 In practice, however, this solution would be unacceptable due to its excessive complexity. In fact, for a message of K M-ary symbols, M K matched filters might be necessary, with about M K comparisons to be performed to select their largest sampled output..5 Sequence Detection To overcome the problems of the last section A. Viterbi of UCLA in 1967 proposed an algorithm for digital communication problems that can be formulated as suitable Marov chains. This algorithm selects the most liely symbol sequence in a number of computations and memory size that grow only linearly with respect to the message length K..5.1 Marov Chains A discrete-time Marov process { Ψ } is a random process that satisfies p Figure -5 (,,...) p( ) Ψ Ψ Ψ = Ψ Ψ. (16) In words, the future sample { Ψ + 1} is independent of past samples Ψ 1, Ψ,... if the present sample is nown. A Marov chain is called homogenous if the conditional probability p ( Ψ Ψ 1) is not a function of. A homogenous Marov chain is therefore a ind of stationary or time invariance. A Marov chain can be visualized by a state transition diagram as shown in Figure -6. Figure -6 Advanced Signal Processing WT 004 Page 9 of 18

10 Another valuable visualization is called the Trellis diagram, first suggested by D. Forney. Figure -7 Figure -7 shows the possible progression of states over time. Each small circle is called a node of the trellis, and corresponds to the Marov chain being a particular state at a particular time. Each arc is called a branch, and corresponds to a particular state transition at a particular time. The branches may be labeled lie the state transitions of the transition diagram. A collection of branches from the beginning to the teral node is called a path..5. Marov Chain Signal Generator Let Ψ be a sequence of a homogenous Marov chain. The sample space of each state Ψ is finite. For the signal generator of interest, the signal samples are a function of the Marov chain state transitions S = g( ψ, ψ + 1) where g(,) a memory less function as in Figure -8. The ML detector has to detect the sequence of states given an observation sequence Y, which is S passed through a noise generator with independent noise components. If h is a finite impulse response, than the inter symbol interference model that was previously developed in (see Figure -3) is an example of a shift-register process, where the observation function is M g( ψ, ψ + ) = ha. (17) 1 i 1 i= 0 The state of the Marov chain is [ X, X,..., X ] Ψ = (18) 1 M Figure -8 where M is the length of the shift-register. Advanced Signal Processing WT 004 Page 10 of 18

11 Example: Consider a discrete-time channel with impulse response h = δ + 0.5δ 1. Using the notation from the shift-register process this is depicted in Figure -9 (a). If the input symbols are independent identically distributed (i.i.d.), the observation Y is a noisy observation of the Marov chain with state diagram (b). Assug binary inputs A, there are only two states in (b) corresponding the two possible previous values for A -1. The arcs are labeled with the input/output (A,S ) of the signal generator. The Trellis diagram loos as follows: Figure -9 Figure The Viterbi Algorithm Every possible path in a trellis has a one-to-one correspondence with an input sequence X,0 K. The goal of a detector, based in the observation S corrupted by noise, is to decide the sequence of inputs. Deciding the sequence of inputs is equivalent to deciding the path through the trellis diagram. Recall that there is a signal s associated with each branch and each stage of the trellis. For each stage there is also an observation y. After observing y, we can assign to each branch of the trellis a numerical value called the branch metric that is low if y is close to s and high otherwise. For the Gaussian case the correct branch metric is branch metric = y s (19) Then for each path through the trellis, the path metric, which is the sum of the branch metrics, is calculated. The preferred path is the one with the lowest path metric. This is nothing else than imizing the Euclidian distance y s ˆ between the received vector y and the decision ŝ, as in section.3. Unfortunately the number of paths through the trellis grows exponentially with the message length K and therefore the computation time and memory usage, too. Advanced Signal Processing WT 004 Page 11 of 18

12 The Viterbi algorithm is a clever way of reducing this exponential growth to linear. Consider one node of the trellis an all paths through the trellis that pass through this node. There are a large number of paths through this node, but all these paths follow just one of four routs routes through this node AB, AD, BC or BD as in Figure -11. Figure -11 The path metric of a path through this node is the sum of the partial path metrics of the portion of the path on the right side of the node and on the left side of the node. Among all the partial paths to the left of the node, the detector pics the one with the smallest partial path metric, which is called the survivor path for that node. If any other partial path would be chosen the overall path metric would be grater than the one with the survivor path, and therefore would not be the smallest path. So every other partial path but the survivor path to the left may be thrown away. The Viterbi algorithm finds the path with the imum path metric by sequentially moving through the trellis and at each node retaining only the survivor paths. At each stage of the trellis it cannot be detered at which node the optimal path passes through, so one survivor Figure -1 path for each and every node must be retained. In a case where there are only two incog branches to a given node at time lie in Figure -1, the two possibilities for the survivor path to the node at time are consisting of the survivor paths at time -1 plus the branches to time. To decide the survivor the partial path metrics for each of those two paths are detered by sumg the partial path metric of the survivor at time -1 and the branch metric. The survivor path at that node is chosen as the path with the smaller partial path metric. Example: The trellis shown below is mared with branch metrics corresponding to the observation sequence {0., 0.6, 0.9, 0.1} of a discrete-time channel with impulse response h = δ + 0.5δ 1 and additive Gaussian noise as in the previous example. The possible state transitions are the following. Figure -13 Advanced Signal Processing WT 004 Page 1 of 18

13 The branches are labeled with their corresponding branch metrics y s. Also the survivor paths for each node and the partial path metrics of each surviving path are shown in Figure -14. Figure -14 A symbol by symbol ML detector would have decided, that the transmitted bits were {0, 1, 1, 0} whereas the ML sequence detector decides for {0, 1, 0, 0} taing into account nowledge of the ISI. The computational complexity of the Viterbi algorithm is the same at each time increment, and hence the total complexity is proportional to the length of K. Yet there is a practical problem remaining. The algorithm does not detere the optimal path until the teral code of the trellis. Since in digital communication systems state sequences may be very long this may result in a very long delay and a very big memory consumption which is generally not affordable. Occasionally all the survivor paths at some time coincide up to some time -d, then the partial paths have merged at depth d, and a decision can be made to all the inputs or states up to time -d. Unfortunately this depends on good fortune and therefore the algorithm is slightly modified to force a decision at time for all transitions prior to time -d, for some truncation depth d. Usually all the partial path metrics of the N partial paths at time are compared and the smallest path is piced. This path represents all the inputs or states until time -d and the node corresponding to this path is chosen to be the input or state at time -d for further computation. If d is chosen to be large enough, this modification will have negligible impact on the probability of detecting the correct sequence. Advanced Signal Processing WT 004 Page 13 of 18

14 3 Error Probability Calculation The sequence error probability is the probability that the path chosen through the trellis does not correspond to the actual state sequence, or in other words the probability that one or more states are in error. As the length of the message K increases this probability usually approaches unity. Fortunately the concern is normally with the probability of a single symbol or bit error, or the probability of a sequence error for a relatively short sequence. This implies that it would be better to optimize for single symbol error, but there are two reasons not to do this. First, it is much more complicated to implement and in fact the best nown algorithms for bitby-bit decision have exponential complexity in K. Secondly, the performance of the ML sequence detector is almost identical at high SNR to the detector that imizes the bit or symbol error probability. 3.1 Error Event It turns out useful to introduce the concept of an error event for the purpose of detering the symbol error probability. Let { Ψ } be the actual state sequence and { Ψˆ } be selected by the Viterbi algorithm. Each error event starts when the actual path and the selected path start to diverge and ends when Figure 3-1 they remerge. The length of an error event is the number of incorrect nodes in the path. So the error event in Figure 3-1 (a) has a length of one and in (b) a length of two. An error event yields in one ore more detection errors, which are incorrect symbols or bits as a result from taing an incorrect path through the trellis. The distance of the error event to the actual path is calculated in the following way. Assume an actual state sequence, and label each branch in the trellis with its squared distance form the corresponding branch of the actual state sequence as in Figure 3-. The actual state sequence will have branch metrics that are zero, and normally all the branches off the actual path sill have a non-zero branch metric. The path metric of each path through the trellis is now the square of the Euclidian distance of that path from the correct all-zero path. Using the branch metric = y s the distance can easily be obtained to be 1.5 for (a) and 3.5 for (b) in Figure 3-. Figure 3- Advanced Signal Processing WT 004 Page 14 of 18

15 3. Symbol Error Probability Now let E denote the set of all error events starting at time i. Each element e of E is characterized by both a correct path Ψ and an incorrect path ˆΨ that diverge and remerge sometime later. Pr[e] is assumed stationary that is Pr [e] is independent of i, the starting time of the error event. The approximation is accurate if the trellis is long relative to the length of the significant error events. Each error event causes one or more detection errors, where a detection error at time means that X at stage of the trellis is incorrect. For the ISI example, each X is a symbol A, so a detection error is the same as a symbol error. Function (0) characterizes the sample times corresponding to detection errors in error event e. 1; if e has a detection error in position m (from the start i) c m = (0) 0; otherwise The probability of a particular error event e starting at time i and causing a detection error at time is c ( e) Pr[ e] i. Of cause, if one error event occurs no other can occur. Therefore the error Events in E are disjoint so i i (1). i= E E i= Pr[detection error at time ] = Pr[ ec ] ( e) = Pr[ e] c ( e) The definition of the total number of detection errors i m () i= m= 0 we () = c () e = c () e yields Pr[detection error] = Pr[ ewe ] ( ). (3) E Hence, the probability of a detection error is equal to the expected number of detection errors in all error events starting at any fixed time i. The probability of the error event Pr[e] is depends on the probabilities of both the correct and incorrect paths Ψ and ˆΨ that mae up e, Pr[ e] = Pr[ Ψ] Pr Ψˆ Ψ. (4) Since it is usually difficult to extract expressions for Pr Ψˆ Ψ, only bounds are derived. Advanced Signal Processing WT 004 Page 15 of 18

16 3..1 Upper Bound of Detection Error Suppose that Ψ is the actual trajectory, corresponding to signals s in Figure 3-3 (a). The region corresponding to the detection of ˆΨ is shown in (b). The probability of the noise carrying the observation in this region is very difficult to calculate, but it is easy to find an upper bound by using a larger decision region as in (c) which ignores the possibility of any trajectory other than Ψ and ˆΨ. For the additive white Gaussian noise model, the probability of region in (c) is Pr ˆ ( ( ˆ Ψ Ψ Q d Ψ, Ψ) / σ ) (5) with Q(.) is defined as 1 α / Qx ( ) = Pr[ X> x] = 1 FX ( x) = e dα π (6) x If Q(x) is defined lie that for a Gaussian random variable X N ( µ, σ ) it becomes x µ Pr[ X > x] = Q σ. (7) Here d( ΨΨ ˆ, ) is the Euclidean distance between transmitted signal ŝ and s corresponding to the state trajectories ˆΨ and Ψ. Hence the probability of a detection error can be written as Pr[detection error] we ( ) Pr[ Ψ] Q( d( Ψˆ, Ψ) / σ ) (8) E If only the set B of error events with imal distance d are considered Pr[detection error] we ( ) Pr[ Ψ ] Q( d / σ ) + other terms (9) E the other terms become insignificant at high SNR. This probability approaches RQ( d / σ ) (30) where R= w()pr[ e Ψ]. (31) B Figure 3-3 Advanced Signal Processing WT 004 Page 16 of 18

17 Example: Here the error event e1 with w(e1)=1 and the distance 1.5 is considered as depicted in Figure 3-4. This is the imum distance. For a trellis with two states there are eight error events where each of them has also w(e)=1. Figure 3-4 So R= w( e) Pr[ Ψ] becomes R = 1Pr[ Ψ]. If all of these error events are equally B liely then R = 1*1/8 = 1 and B B ( σ ) Pr[detection error] Q 1.5 / + other terms. E Figure Lower Bound of Detection Error To obtain the lower bound the decision region of (d) can be used. This is the decision region where any error event conditioned on the actual state sequence Ψ is shown. The probability of any error event is clearly lower bounded by calculating the smaller decision region shown in (c). Figure 3-6 Since we () 1for all error events e the probability of a detection error can be written as Pr[detection error] Pr[ e] = Pr[an error event]. (3) E For one particular path through the trellis let d ( ) Ψ denote the distance of the imum distance error event. Let d be the imum distance error event over all possible actual state sequences Ψ. Then d ( Ψ) d. To mae the bound strongest, obviously the closest error event to Ψ has to be chosen, which is one of those at distance d ( Ψ ). Than the probability of any error event becomes Pr[an error event Ψ] Qd ( ( Ψ ) / σ ). (33) Advanced Signal Processing WT 004 Page 17 of 18

18 Consequently the probability of a detection error can be written as Pr[detection error Ψ] Qd ( ( Ψ ) / σ ). (34) And using the total probability this is Pr[detection error] Pr[ Ψ] Qd ( ( Ψ) / σ ). (35) Ψ It is save to omit some terms here because all terms are positive. So consider only those state sequences Ψ for which d ( Ψ ) = d. Pr[detection error] Pr[ Ψ] Qd ( ( Ψ) / σ ) (36) Ψε A Here A is the set of actual paths Ψ that have the imum distance error event, and d is that imum distance. Define the probability that a randomly chosen Ψ has an error event starting at any fixed time with distance d as P = Pr[ Ψ]. (37) Ψε A Than the lower bound can be expressed as Pr[detection error] PQ( d ( Ψ ) / σ ). (38) 3..3 Upper and Lower Bounds of Detection Error Combining upper an lower bounds, PQ( d / σ ) Pr[detection error] RQ( d ( Ψ ) / σ ) (39), where the upper bound is approximate since some terms were thrown away. To conclude, at high SNR Pr[detection error] CQ( d / σ ) for some constant C between P and R. (40) Example: From the previous example note that P=R=1. Hence Pr[detection error] Q( 1.5 / σ ). Here each detection error causes exactly one bit error, so Pr[detection error] = Pr[bit error]. Hence, with this sequence detector approximately the same probability of error is obtained as for an isolated pulse and a matched filter receiver. In general a single detection error may cause more than one bit error. Suppose each input of the Marov chain X is detered by n source bits (hence the alphabet has size n ). Then each detection error causes at least one and at most n bit errors. 1 Pr[detection error] Pr[bit error] Pr[detection error] (41) n The typical assumption is pessimistic: Pr[detection error] Pr[bit error] (4) Advanced Signal Processing WT 004 Page 18 of 18